Prefill and decode want different expert maps
·4 min read
On one DGX Station, a small change to where DeepSeek-V4.1-Flash keeps its experts made reading a 16,000-token prompt 7% faster. Decode speed did not change, and long-context accuracy held. We measured
On one DGX Station, a small change to where DeepSeek-V4.1-Flash keeps its experts made reading a 16,000-token prompt 7% faster. Decode speed did not change, and long-context accuracy held. We measured it twice, in two separate sessions: 1.070 and 1.075 times faster. Pushing the same idea further made prompt reading up to 34% faster, but then 16-user decode got 4 to 9% slower. Reading a prompt and writing an answer want the experts in different places. This is how we found that out, and where we
8
0
0
0